Mercury 2.5 Preview vs Step 5 Preview
Inception Labs · US | StepFun · China · Updated June 2026
Quick verdict
Pick Mercury 2.5 Preview for very high inference throughput (1,107 tok/s) via diffusion-based parallel token generation or coding accuracy (95.7%, 91st percentile among cost-optimized models). Pick Step 5 Preview for sparse moe architecture — 600b total parameters, only 27b active per token — built for long-horizon agentic work across large codebases and documents or 1m-token context window (64k max output) at $1/$2.70 per million tokens, among the cheapest frontier-scale context windows available. On a tight budget at scale, Mercury 2.5 Preview is the value pick.
Mercury 2.5 Preview (Inception Labs, US) and Step 5 Preview (StepFun, China) line up two different AI ecosystems against each other — a comparison that is as much about cost philosophy and openness as raw capability. Mercury 2.5 Preview is inception Labs' August 31, 2026 diffusion-based language model preview, refining tokens in parallel for roughly 10x the throughput of comparable autoregressive models at cost-optimized-tier quality. Step 5 Preview is stepFun's September 20, 2026 flagship preview — a 600B-parameter sparse MoE model (27B active) with a 1M-token context window at $1/$2.70 per million tokens, open weights promised for mid-October. They diverge most on price and context window — each quantified below from the models' real specs.
Key differences at a glance
- ▸Price: Mercury 2.5 Preview is about 25× cheaper on input ($0.04/$0.15 per 1M tokens vs $1/$2.7 per 1M tokens) — a large enough gap that at scale it can be the single biggest line item in the decision.
- ▸Context window: Step 5 Preview holds 3.8× more — 1M tokens (~1,500 pages) vs 260K tokens (~390 pages). But effective recall usually fades long before the advertised ceiling, so the bigger number only helps if the model reasons over it.
- ▸Recency: Step 5 Preview is the newer model by about 20 days (released September 20, 2026), usually meaning fresher training data and capabilities.
- ▸Ecosystem: this is a US-vs-China matchup — they differ in pricing philosophy, data-residency options, and tooling ecosystems, not only benchmarks.
Side-by-side specs
| Spec | Mercury 2.5 Preview | Step 5 Preview |
|---|---|---|
| Provider | Inception Labs (US) | StepFun (China) |
| Released | August 31, 2026 | September 20, 2026 |
| Context window | 260K tokens (~390 pages) | 1M tokens (~1,500 pages) |
| Price (in/out) | $0.04/$0.15 per 1M tokens | $1/$2.7 per 1M tokens |
| Open weight? | No — API only | No — API only |
| Modalities | text | text, image, code |
| SWE-Bench Verified | Not published | Not published |
| MRCR v2 @ 1M | Not published | Not published |
Who wins what
Very high inference throughput (1,107 tok/s) via diffusion-based parallel token generation
Mercury 2.5 Preview
Inception Labs' August 31, 2026 diffusion-based language model preview, refining tokens in parallel for roughly 10x the throughput of comparable autoregressive models at cost-optimized-tier quality — and it runs cheaper at $0.04/$0.15 per 1M tokens.
Coding accuracy (95.7%, 91st percentile among cost-optimized models)
Mercury 2.5 Preview
At $0.04/$0.15 per 1M tokens it undercuts Step 5 Preview ($1/$2.7 per 1M tokens), and that gap compounds at volume.
Mathematics accuracy (97.0%, 97th percentile)
Mercury 2.5 Preview
Mercury 2.5 Preview lists mathematics accuracy (97.0%, 97th percentile) among its strengths; Step 5 Preview does not.
Sparse MoE architecture — 600B total parameters, only 27B active per token — built for long-horizon agentic work across large codebases and documents
Step 5 Preview
Its 1M tokens window holds about 3.8× more than Mercury 2.5 Preview's 260K tokens in a single prompt.
1M-token context window (64K max output) at $1/$2.70 per million tokens, among the cheapest frontier-scale context windows available
Step 5 Preview
Mercury 2.5 Preview is comparatively weak here — 260K context window is far shorter than frontier 1M-token models
67.7% on DeepSWE v1.1, ahead of open-weight rivals Kimi K3 (67.5%) and GLM-5.3 (66.9%)
Step 5 Preview
StepFun's September 20, 2026 flagship preview — a 600B-parameter sparse MoE model (27B active) with a 1M-token context window at $1/$2.70 per million tokens, open weights promised for mid-October — and it carries the larger 1M tokens context.
Lowest cost at scale
Mercury 2.5 Preview
At $0.04/$0.15 per 1M tokens, it is the cheaper of the two — the gap dominates the bill on high-volume workloads.
Largest single-prompt input
Step 5 Preview
Its 1M tokens window is about 3.8× larger than Mercury 2.5 Preview's 260K tokens, fitting roughly 1,500 pages in one prompt.
Which should you pick?
A cost-sensitive startup shipping high volume
→ Mercury 2.5 Preview
At $0.04/$0.15 per 1M tokens it undercuts Step 5 Preview, and on millions of tokens that margin decides the monthly bill.
Someone analysing very long documents or codebases
→ Step 5 Preview
Larger 1M tokens window fits more in one prompt.
Anyone whose priority is very high inference throughput (1,107 tok/s) via diffusion-based parallel token generation
→ Mercury 2.5 Preview
It is specifically built for that.
Anyone whose priority is sparse moe architecture — 600b total parameters, only 27b active per token — built for long-horizon agentic work across large codebases and documents
→ Step 5 Preview
That is its strongest area.
An enterprise with regional data-residency rules
→ Mercury 2.5 Preview or Step 5 Preview
Origin (US vs China) affects where data is processed and which compliance regime applies — check the provider's terms for your region.
Mercury 2.5 Preview: where it fits
Inception Labs' August 31, 2026 diffusion-based language model preview, refining tokens in parallel for roughly 10x the throughput of comparable autoregressive models at cost-optimized-tier quality. Released August 31, 2026 by Inception Labs, it is built for very high inference throughput (1,107 tok/s) via diffusion-based parallel token generation, coding accuracy (95.7%, 91st percentile among cost-optimized models), mathematics accuracy (97.0%, 97th percentile), and tunable reasoning levels with parallel tool calls and schema-aligned JSON output.
Its trade-offs are real: 260K context window is far shorter than frontier 1M-token models, no vision or audio modalities, positioned only against cost-optimized models (GPT-5.6 Luna Low, Gemini 3.5 Flash-Lite), not frontier-class, and current $0.04/$0.15 pricing includes a limited-time launch promotion on OpenRouter (list price is $0.20/$0.75, may rise after Sept 8, 2026). At $0.04 in / $0.15 out per million tokens, it sits in the budget price band.
Step 5 Preview: where it fits
StepFun's September 20, 2026 flagship preview — a 600B-parameter sparse MoE model (27B active) with a 1M-token context window at $1/$2.70 per million tokens, open weights promised for mid-October. Released September 20, 2026 by StepFun, it is built for sparse MoE architecture — 600B total parameters, only 27B active per token — built for long-horizon agentic work across large codebases and documents, 1M-token context window (64K max output) at $1/$2.70 per million tokens, among the cheapest frontier-scale context windows available, 67.7% on DeepSWE v1.1, ahead of open-weight rivals Kimi K3 (67.5%) and GLM-5.3 (66.9%), and scores 44 on the Artificial Analysis Intelligence Index, ranking 27th of 653 models tracked.
Its trade-offs: trails closed frontier models on coding — 67.7% on DeepSWE v1.1 vs GPT-6 Astra's 74.1% and Claude Opus 5's 74.0%, open weights were promised for October 15, 2026 but hadn't shipped as of this writing — API-only for now despite the open-weight roadmap, a 'Preview' release, not StepFun's finished flagship — positioning and pricing may change once weights ship, and stepCodeBench and DeepSWE figures are StepFun's own reported numbers, not yet independently reproduced at scale. At $1 in / $2.7 out per million tokens, it sits in the budget price band.
The bottom line for this matchup
This is less "which is smarter" and more "which ecosystem fits." Mercury 2.5 Preview (US) and Step 5 Preview (China) differ on pricing philosophy, data-residency, and tooling as much as on raw scores. Mercury 2.5 Preview is the cheaper option, which matters at volume. The pragmatic move is to run one real task through both and judge the outputs against your own constraints — including where your data is allowed to be processed.
Want both Mercury 2.5 Preview and Step 5 Preview without two subscriptions? LumiChats gives you these plus 40+ models under one ₹69/day pass (about $1/day) — draft with one, cross-check with the other.
See pricingFrequently asked questions
Is Mercury 2.5 Preview or Step 5 Preview better for coding?
Public SWE-Bench figures are not available for either model, so the honest test is your own repository — run an identical real bug through both. By design, Mercury 2.5 Preview leans toward very high inference throughput (1,107 tok/s) via diffusion-based parallel token generation while Step 5 Preview leans toward sparse moe architecture — 600b total parameters, only 27b active per token — built for long-horizon agentic work across large codebases and documents, and that positioning usually predicts which feels better on your codebase.
Which is cheaper, Mercury 2.5 Preview or Step 5 Preview?
Mercury 2.5 Preview is cheaper — $0.04/$0.15 per 1M tokens vs $1/$2.7 per 1M tokens, roughly 25× apart on input.
Which has the bigger context window?
Step 5 Preview — 1M tokens vs 260K tokens, about 3.8× larger. Useful only if the model actually reasons over the full window, which not all do.
Can I use both Mercury 2.5 Preview and Step 5 Preview together?
Yes — a multi-model platform like LumiChats gives you Mercury 2.5 Preview, Step 5 Preview and 40+ others under one ₹69/day pass (about $1/day), so you can draft with one and cross-check with the other instead of buying two subscriptions.
Which is newer, Mercury 2.5 Preview or Step 5 Preview?
Step 5 Preview — released September 20, 2026, about 20 days after Mercury 2.5 Preview.
Related comparisons
Specifications and benchmarks reflect publicly reported figures as of June 2026 and may change as providers release updates. Always verify on your own workload.